[15748] | 1 | #region License Information
|
---|
| 2 |
|
---|
| 3 | /* HeuristicLab
|
---|
| 4 | * Copyright (C) 2002-2018 Heuristic and Evolutionary Algorithms Laboratory (HEAL)
|
---|
| 5 | *
|
---|
| 6 | * This file is part of HeuristicLab.
|
---|
| 7 | *
|
---|
| 8 | * HeuristicLab is free software: you can redistribute it and/or modify
|
---|
| 9 | * it under the terms of the GNU General Public License as published by
|
---|
| 10 | * the Free Software Foundation, either version 3 of the License, or
|
---|
| 11 | * (at your option) any later version.
|
---|
| 12 | *
|
---|
| 13 | * HeuristicLab is distributed in the hope that it will be useful,
|
---|
| 14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
---|
| 15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
---|
| 16 | * GNU General Public License for more details.
|
---|
| 17 | *
|
---|
| 18 | * You should have received a copy of the GNU General Public License
|
---|
| 19 | * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
|
---|
| 20 | */
|
---|
| 21 |
|
---|
| 22 | #endregion
|
---|
| 23 |
|
---|
| 24 | using System;
|
---|
| 25 | using System.Collections.Generic;
|
---|
| 26 | using System.Linq;
|
---|
| 27 | using HeuristicLab.Core;
|
---|
| 28 |
|
---|
| 29 | namespace HeuristicLab.Random {
|
---|
| 30 | /// <summary>
|
---|
| 31 | /// See Radford Neal, MCMC using Hamiltonian dynamics, 2011, https://arxiv.org/pdf/1206.1901.pdf
|
---|
| 32 | /// Algorithm from the paper:
|
---|
| 33 | ///
|
---|
| 34 | /// HMC = function (U, grad_U, epsilon, L, current_q)
|
---|
| 35 | /// {
|
---|
| 36 | /// q = current_q
|
---|
| 37 | /// p = rnorm(length(q), 0, 1) # independent standard normal variates
|
---|
| 38 | /// current_p = p
|
---|
| 39 | /// # Make a half step for momentum at the beginning
|
---|
| 40 | /// p = p - epsilon * grad_U(q) / 2
|
---|
| 41 | /// # Alternate full steps for position and momentum
|
---|
| 42 | /// for (i in 1:L)
|
---|
| 43 | /// {
|
---|
| 44 | /// # Make a full step for the position
|
---|
| 45 | /// q = q + epsilon* p
|
---|
| 46 | /// # Make a full step for the momentum, except at end of trajectory
|
---|
| 47 | /// if (i!=L) p = p - epsilon * grad_U(q)
|
---|
| 48 | /// }
|
---|
| 49 | /// # Make a half step for momentum at the end.
|
---|
| 50 | /// p = p - epsilon* grad_U(q) / 2
|
---|
| 51 | /// # Negate momentum at end of trajectory to make the proposal symmetric
|
---|
| 52 | /// p = -p
|
---|
| 53 | /// # Evaluate potential and kinetic energies at start and end of trajectory
|
---|
| 54 | /// current_U = U(current_q)
|
---|
| 55 | /// current_K = sum(current_p^2) / 2
|
---|
| 56 | /// proposed_U = U(q)
|
---|
| 57 | /// proposed_K = sum(p^2) / 2
|
---|
| 58 | /// # Accept or reject the state at end of trajectory, returning either
|
---|
| 59 | /// # the position at the end of the trajectory or the initial position
|
---|
| 60 | /// if (runif(1) < exp(current_U-proposed_U+current_K-proposed_K)) {
|
---|
| 61 | /// return (q) # accept
|
---|
| 62 | /// }
|
---|
| 63 | /// else {
|
---|
| 64 | /// return (current_q) # reject
|
---|
| 65 | /// }
|
---|
[15749] | 66 | /// }
|
---|
[15748] | 67 |
|
---|
| 68 | /// </summary>
|
---|
| 69 | public static class HamiltonianMonteCarlo {
|
---|
| 70 |
|
---|
| 71 | public static IEnumerable<double[]> SampleChain(double[] startingPosition,
|
---|
| 72 | Func<double[], Tuple<double, double[]>> potentialEnergyFunction,
|
---|
| 73 | IRandom uniformRandom, double stepSize, int steps = 10) {
|
---|
| 74 | var x = (double[])startingPosition.Clone();
|
---|
| 75 | while(true) {
|
---|
| 76 | x = Sample(x, potentialEnergyFunction, uniformRandom, stepSize, steps);
|
---|
| 77 | yield return x;
|
---|
| 78 | }
|
---|
| 79 | }
|
---|
| 80 |
|
---|
| 81 | public static double[] Sample(
|
---|
| 82 | double[] startingPosition,
|
---|
| 83 | Func<double[], Tuple<double, double[]>> potentialEnergyFunction, // U
|
---|
| 84 | IRandom uniformRandom, double stepSize, int steps) {
|
---|
| 85 | // step 1
|
---|
| 86 | // new values for the momentum variables are randomly drawn from their Gaussian
|
---|
| 87 | // distribution, independently if the current values of the position variables
|
---|
| 88 |
|
---|
| 89 | // step 2
|
---|
| 90 | // Metropolis update is performed, using Hamiltonian dynamics to propose a new state.
|
---|
| 91 | // Starting with the current state Hamiltonian dynamics is simulated for L steps
|
---|
| 92 | // using the Leapfrog method with a step size of eps.
|
---|
| 93 | // L and eps are parameters of the algorithm, which need to be tuned to obtain
|
---|
| 94 | // good performance.
|
---|
| 95 | // The proposed state is accepted as the next state of the Markov chain with
|
---|
| 96 | // probability [...].
|
---|
| 97 | // If the proposed state is not accepted the next state is the same as the current state
|
---|
| 98 | // (and is counted again when estimating the expectation of some function of
|
---|
| 99 | // state by its average over states of the Markov chain).
|
---|
| 100 |
|
---|
| 101 |
|
---|
| 102 | int nVars = startingPosition.Length;
|
---|
| 103 | var u_initial = potentialEnergyFunction(startingPosition).Item1;
|
---|
| 104 |
|
---|
| 105 | // rename variables to match Neal's algorithm
|
---|
| 106 | var current_q = startingPosition;
|
---|
| 107 | var epsilon = stepSize;
|
---|
| 108 | var L = steps;
|
---|
| 109 |
|
---|
| 110 | var q = current_q;
|
---|
| 111 | // p = rnorm(length(q), 0, 1) # independent standard normal variates
|
---|
| 112 | var p = q
|
---|
| 113 | .Select(_ => NormalDistributedRandom.NextDouble(uniformRandom, 0, 1))
|
---|
| 114 | .ToArray();
|
---|
| 115 | var current_p = p;
|
---|
| 116 | // Make a half step for momentum at the beginning
|
---|
| 117 | // p = p - epsilon * grad_U(q) / 2;
|
---|
| 118 | p = VectorSum(p, ScaleVector(-0.5 * epsilon, potentialEnergyFunction(q).Item2));
|
---|
| 119 | // Alternate full steps for position and momentum
|
---|
| 120 |
|
---|
| 121 | for (int i = 1; i <= L; i++) {
|
---|
| 122 | // Make a full step for the position
|
---|
| 123 | // q = q + epsilon * p;
|
---|
| 124 | q = VectorSum(q, ScaleVector(epsilon, p));
|
---|
| 125 | // Make a full step for the momentum, except at end of trajectory
|
---|
| 126 | if (i != L)
|
---|
| 127 | p = VectorSum(p, ScaleVector(-epsilon, potentialEnergyFunction(q).Item2));
|
---|
| 128 | }
|
---|
| 129 | // Make a half step for momentum at the end.
|
---|
| 130 | /// p = p - epsilon* grad_U(q) / 2
|
---|
| 131 | p = VectorSum(p, ScaleVector(-0.5 * epsilon, potentialEnergyFunction(q).Item2));
|
---|
| 132 | // Negate momentum at end of trajectory to make the proposal symmetric
|
---|
| 133 | /// p = -p
|
---|
| 134 | p = ScaleVector(-1.0, p);
|
---|
| 135 | // Evaluate potential and kinetic energies at start and end of trajectory
|
---|
| 136 | double current_U = potentialEnergyFunction(current_q).Item1;
|
---|
| 137 | // sum(current_p ^ 2) / 2
|
---|
| 138 | double current_K = 0.5 * current_p.Sum(pi => pi * pi);
|
---|
| 139 | double proposed_U = potentialEnergyFunction(q).Item1;
|
---|
| 140 | // sum(p^2) / 2
|
---|
| 141 | double proposed_K = 0.5 * p.Sum(pi => pi * pi);
|
---|
| 142 | /// # Accept or reject the state at end of trajectory, returning either
|
---|
| 143 | /// # the position at the end of the trajectory or the initial position
|
---|
| 144 | if (uniformRandom.NextDouble() < Math.Exp(current_U - proposed_U + current_K - proposed_K)) {
|
---|
| 145 | return q; // accept
|
---|
| 146 | } else {
|
---|
| 147 | return current_q; // # reject
|
---|
| 148 | }
|
---|
| 149 | }
|
---|
| 150 |
|
---|
| 151 | #region to be improved or removed
|
---|
| 152 | private static double VectorProd(double[] x, double[] y) {
|
---|
| 153 | double s = 0.0;
|
---|
| 154 | for (int i = 0; i < x.Length; i++)
|
---|
| 155 | s += x[i] * y[i];
|
---|
| 156 | return s;
|
---|
| 157 | }
|
---|
| 158 |
|
---|
| 159 | private static double[] VectorSum(double[] x, double[] y) {
|
---|
| 160 | double[] s = new double[x.Length];
|
---|
| 161 | for (int i = 0; i < x.Length; i++)
|
---|
| 162 | s[i] += x[i] + y[i];
|
---|
| 163 | return s;
|
---|
| 164 | }
|
---|
| 165 |
|
---|
| 166 | private static double[] ScaleVector(double a, double[] x) {
|
---|
| 167 | double[] s = new double[x.Length];
|
---|
| 168 | for (int i = 0; i < x.Length; i++)
|
---|
| 169 | s[i] = a * x[i];
|
---|
| 170 | return s;
|
---|
| 171 | }
|
---|
| 172 | #endregion
|
---|
| 173 | }
|
---|
| 174 | }
|
---|